How this data is made
Counted on 2026-10-18 from our working records (input stamp a2785546a2e8)
This page says which steps of our work an AI performs, which checks are made by a person, and how many of each we can count. Every number below is counted by a script from the record named beside it; none is typed. Next to each number is the rule the script used to count it.
Where AI is used
- AI assistants draft, edit and check our records, pages and documents, including this page. Our pages carry the line "AI-assisted, human-verified"; the sections below say what that does and does not cover.
- Our work is organised in written directives. A directive's receipt file records what was measured and what was changed.
- Recent corrections to records are made by scripts built on a shared writer that keeps the value it replaces in the record and reads the record back after writing. Older records were written by earlier scripts.
- Some changes are reviewed by a second AI session before they reach the public sites. Our records call this a "section 6 co-sign".
What people check, and what AI checks, counted apart
The two kinds of check are never added together. A review by an AI session is never counted as a human check.
| Kind | What is counted | Count | Rule used to count it |
|---|---|---|---|
| Human | Operator decisions and actions recorded in the operator actions log | 331 | Distinct entries whose heading is "R-" followed by a number, in tracking/operator-actions-log.md. The log is kept by our AI sessions from what the operator reports or does. |
| Human | Deployments run by the operator | 163 | Rows of tracking/deploy-log.md whose last cell (the "By" column) names the operator. 10 rows of the log carry a stray separator in an earlier cell; they are read by their last cell, not by position. |
| AI | Reviews by a second AI session recorded as granted | 257 | Rows of tracking/DIRECTIVES-STATUS.json (890 rows) whose section6_state column is "granted". The free-text column is not read. |
| AI | Reviews recorded as partly granted (not included in the line above) | 23 | Rows whose section6_state column is "partial". |
| AI | Reviews recorded as required and not yet given (not included above) | 5 | Rows whose section6_state column is "required": the row says a second review is needed before a change to what we publish. The free text of 5 of these 5 rows begins with "required"; the count does not say a review has been asked for. |
| AI | Reviews recorded as still pending (not included above) | 0 | Rows whose section6_state column is "pending": a review is recorded as awaited. The free text of 0 of these 0 rows begins with "pending"; the rest begin differently. "Required" and "pending" are counted apart and neither is in the granted line. |
| AI | Rows with no review state, whose free text reads like a co-sign (not included above) | 23 | Rows whose section6_state column is empty and whose free-text column says a co-sign was given. The count does not say the review was granted. |
| AI | Rows with no review state and no co-sign text (not included above) | 11 | Rows whose section6_state column is empty and whose free text is not a co-sign. The count does not say whether a second review applied to them. |
The measured error record
Our public corrections page listed 67 open corrections when this page was built. That is the number printed on that page, not a count of every error.
Limits
- No qualified toxicologist or risk assessor has reviewed our method.
- 4 findings filed under the kind "method" in our findings log are open; the log lists every open finding, under whatever kind it was filed. The kind is a filing category, not a subject: this is a count of that category and says nothing about how many findings concern the method that turns records into risk bands.
- A count of reviews says how many times a review step was recorded. It does not say that every record has been checked by a person.